Negative sentiment share: a better metric than average sentiment

What negative sentiment share is

Negative sentiment share answers one question per topic: when customers mention this, how often is it a complaint? If 340 reviews mention billing and 210 of those mentions are negative, billing's negative share is 62%.

It is a proportion, so it is comparable across stores, months and topics of very different sizes, and anyone can check it by counting.

Why not an average sentiment score?

Many tools give each review or topic a sentiment score, for example from minus one to plus one, and average it. Averages have two problems for customer feedback.

• They hide polarisation. A topic where half the customers are delighted and half furious averages to neutral, which describes nobody.
• They are hard to interpret. Is 0.18 good? Is a drop from 0.21 to 0.18 meaningful? Nobody outside the analytics team can say.

Negative share does not have either problem. A topic at 50% negative is clearly contested, and a rise from 30% to 45% is clearly a change anyone can understand.

How to calculate it

1. Count mentions, not reviews. "Great staff, slow billing" is one positive staff mention and one negative billing mention.
2. For each topic, divide negative mentions by all mentions with a known sentiment.
3. Report unknown or unclear sentiment separately, rather than counting it as neutral or dropping it silently.
4. If one review mentions the same topic positively and negatively, decide on a rule (for example, count it once as negative) and apply it consistently.
5. Show the number of mentions next to every share.

Two different shares

Be clear which one you report. Negative share within a topic (negative mentions of billing divided by all billing mentions) says how bad billing is when it comes up. Prevalence (reviews with a negative billing mention divided by all reviews) says how much of your total feedback is billing complaints. Both are useful; mixing them up causes confusion.

• Within-topic negative share: best for judging the experience of a topic.
• Prevalence of complaints: best for judging the size of the problem across all customers.

Reading it well

• With volume. 80% negative on 5 mentions is not the same as 80% on 500.
• With trend. Compare with the previous period for the same stores and sources.
• With context. Some topics are naturally more negative: customers mention parking mostly to complain. Compare a topic with itself over time and with peer stores, not with other topics.
• With the reviews. Read a sample before acting on any change.

Pitfalls

• Small samples swing wildly. Set a minimum before ranking.
• Changing classification rules changes the share. Note when the taxonomy or model changed.
• Mixing sources. App reviews and store reviews describe different experiences; report them separately or say clearly when combined.
• Treating neutral as positive. Keep unknown sentiment visible.

A worked example

Two stores each have 200 reviews that mention staff.

• Store A: 150 positive mentions, 50 negative. Negative share 25%.
• Store B: 100 positive, 100 negative. Negative share 50%.

On a scale of minus one to plus one where each mention counts fully, Store A's average is plus 0.5 and Store B's is 0. Both numbers say A is better, but only the negative share says how: at Store B, one customer in two who mentions staff is complaining. That is a sentence a store manager can act on.

Now suppose Store B's share was 30% three months ago. The rise from 30% to 50% on a steady 200 mentions is a clear change. An average sentiment that moved from 0.4 to 0 says the same thing far less clearly.

Its partner: positive share

The same calculation for positive mentions shows strengths. A topic with high positive share and high volume is something customers value: patient staff, a good exchange policy, quick service. Track it too. When a strength's positive share starts to fall, it is often the earliest sign of a problem, long before complaints rise.

Where to use it

• Monthly reporting by topic, for the whole network and per store.
• Store comparisons within peer groups.
• Before-and-after checks on fixes.
• Prioritisation, combined with volume and trend.

Doing this with Falcon

Falcon reports negative share for every topic and month in its experience signals, counted from the theme mentions in each review rather than the review as a whole, with unknown sentiment kept visible and small samples flagged. Every share opens the reviews behind it.